In Silico Prediction Tools for Cytochrome P450 Isoform Specificity: A Comprehensive Review and Model Evaluation

In humans, cytochrome P450 (CYP450) enzymes play a pivotal role in catalyzing over 90% of the enzymatic reactions associated with xenobiotic metabolism. Accurately predicting whether interested chemicals act as substrates or inhibitors of different CYP450 isoforms can assist in preselecting hit compounds and optimizing lead compounds for further drug discovery and chemical toxicology study. This work embarks on a comprehensive overview of recently developed in silico prediction tools of CYP450 isoform specificity. We summarize information on these models by two major categories of computational approaches: structure-based and ligand-based. We also analyze various aspects of these models, including their datasets, algorithms, and performance metrics. Subsequently, we employ 100 of the most frequently prescribed drugs to evaluate ten prediction tools which were developed via different classical machine learning methods (e.g., support vector machine, random forest, learning-based approaches, etc.) or deep learning approaches (e.g., graph attention neural network and ESM-1b Transformer). We point out that deep learning-based models like admetLab and ESP can quickly and accurately predict results. However, the coverage and performance of the investigated models are constrained by the limited quantity and quality of their datasets. We discuss both the advantages and limitations of the evaluated models, providing guidance for selecting appropriate computational tools to carry out predictions, and highlighting trends in the field of computational CYP450 isoform specificity prediction.